[XLA:GPU] Add support for NCCL ncclCommInitRankScalable API #21273
+171
−40
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
nvcastet
force-pushed
the
ncclCommInitRankScalable
branch
from
40286ab to
d2fb81e
Compare
|
CC @ezhulenev |
nvcastet
force-pushed
the
ncclCommInitRankScalable
branch
3 times, most recently
from
8a46f76 to
86e50af
Compare
nvcastet
force-pushed
the
ncclCommInitRankScalable
branch
from
86e50af to
98ef02d
Compare
ezhulenev
reviewed
Jan 10, 2025
ezhulenev
reviewed
Jan 10, 2025
ezhulenev
reviewed
Jan 10, 2025
ezhulenev
reviewed
Jan 10, 2025
ezhulenev
reviewed
Jan 10, 2025
ezhulenev
reviewed
Jan 10, 2025
| // Returns true if this clique is a subset of `other`: both cliques have the | ||
| // same `stream_id` and all clique devices are part of `other` clique. | ||
| bool IsSubsetOf(const CliqueKey& other) const final; | ||
|
|
||
| // Returns a copy of the key (subkey) with the root device properly set given | ||
| // nroots and root_seq_id. The subkey is used to generate a NcclCliqueId. |
There was a problem hiding this comment.
I feel that a bit more details would be helpful, I don't have the context, but the prorblem is that after reading this documentation I still don't have any idea what's going on :) (cc @frgossen)
There was a problem hiding this comment.
I tried to be a bit more verbose. Let me know if it is better.
The NCCL doc is at https://docs.nvidia.com/deeplearning/nccl/user-guide/docs/usage/communicators.html#creating-a-communicator-using-multiple-nccluniqueids
nvcastet
force-pushed
the
ncclCommInitRankScalable
branch
3 times, most recently
from
06fcca9 to
bb14e02
Compare
nvcastet
force-pushed
the
ncclCommInitRankScalable
branch
from
bb14e02 to
f146a48
Compare
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
ncclCommInitRankScalableenables the initialization of communicators via multiple roots which improves the init performance at large scale.The maximum number of ranks associated with a root rank to initialize a NCCL communicator can be tuned via
--xla_gpu_nccl_init_max_rank_per_root_ratio. Default is 128 ranks per root.